Learning to Adapt to Evolving Domains

NeurIPS 2020 11-16-2020

Presentation

slides could be found from the following link:

15MB

Learning to adapt to evolving domains-slides.pptx

presentation slides by Changbin Li

Questions

Q1: What is the role of Theorem 1?

1) approximate the loss function because no access to the label of target data

Q2: What is the fundamental difference between EAML and MAML?

MAML: learn a good initialization

EAML: learn meta-representation and meta-adapter

Q3: What is the fundamental difference between the setting in the paper and continual learning?

Q4: What is A-distance? How could alpha affect the model?

Q5: Different assumptions about domain adaptation and domain generalization?

Q6: Catastrophic forgetting?

Reference

• https://papers.nips.cc/paper/2020/file/fd69dbe29f156a7ef876a40a94f65599-Paper.pdf

• https://github.com/Liuhong99/EAML